Evaluation in early visual processing – evidence from deep neural networks

An important question in affective science is whether valence assignment, i.e., evaluating a perceived object as positive or negative, occurs after perception or whether it is intrinsic to it. Recent theories propose that valence assignment, or simply evaluation, may be embedded within conscious perception. To test this idea, we combined deep neural network (DNN) modeling with representational similarity analysis (RSA) to identify which levels of visual processing contribute to valence assignment. Participants performed a forced-choice preference task, selecting one of three emotionally neutral everyday objects (e.g., cups, lamps) under varying instructions: no instruction, utility-based, aesthetic-based, and under time pressure. Perceived valence distances between objects were captured as representational dissimilarity matrices (RDMs) and compared with DNN activation distances across layers. Across studies, valence distances correlated with differences in the first DNN layer, reflecting low-level visual processing. However, this relationship could not be explained by individual features such as contrast or luminance, implying a more integrative perceptual basis. When choices involved objects from distinct categories, valence judgments primarily reflected perceived utility, yet aesthetic instructions biased them toward similar visual computations. These findings indicate that valence assignment is partially built into perceptual processing, with higher-level task demands flexibly shaping its expression. This supports a view of conscious perception as inherently valenced, yet context-sensitive.